Theory-guided deep-learning for electrical load forecasting (TgDLF) via ensemble long short-term memory
Yuntian Chen, Dongxiao Zhang
Abstract
Electricity constitutes an indispensable source of secondary energy in modern society. Accurate and robust short-term electrical load forecasting is essential for more effective scheduling of load generation, minimizing the gap between generation and demand, and reducing electricity losses. This study proposes theory-guided deep-learning load forecasting (TgDLF), which is a gradient-free model that fully combines domain knowledge and machine learning algorithms. TgDLF predicts the future load through load ratio decomposition, in which dimensionless trends are obtained based on domain knowledge, and the local fluctuations are estimated via data-driven models. TgDLF simplifies the problem with the assistance of expertise, and utilizes the strong expressive power of neural networks to obtain accurate predictions. The historical load, weather forecast and calendar effect are considered in the model, and the model's robustness to inaccurate weather forecast data is improved by adding synthetic disturbance during the training process. Cross-validation experiments demonstrate that TgDLF is 23% more accurate than long short-term memory, and the TgDLF with enhanced robustness can effectively extract information from weather forecast data with up to 40% noise.